Limits...
Broad-Band Visually Evoked Potentials: Re(con)volution in Brain-Computer Interfacing.

Thielen J, van den Broek P, Farquhar J, Desain P - PLoS ONE (2015)

Bottom Line: We defined a linear generative model that decomposes full responses into overlapping single-flash responses.In an online experiment, 12 participants tested a 6 × 6 matrix speller BCI.These predicted responses are proven to be well-suited as templates for a BBVEP-based BCI, thereby enabling communication and control by brain activity only.

View Article: PubMed Central - PubMed

Affiliation: Radboud University Nijmegen, Donders Center for Cognition, Nijmegen, Netherlands.

ABSTRACT
Brain-Computer Interfaces (BCIs) allow users to control devices and communicate by using brain activity only. BCIs based on broad-band visual stimulation can outperform BCIs using other stimulation paradigms. Visual stimulation with pseudo-random bit-sequences evokes specific Broad-Band Visually Evoked Potentials (BBVEPs) that can be reliably used in BCI for high-speed communication in speller applications. In this study, we report a novel paradigm for a BBVEP-based BCI that utilizes a generative framework to predict responses to broad-band stimulation sequences. In this study we designed a BBVEP-based BCI using modulated Gold codes to mark cells in a visual speller BCI. We defined a linear generative model that decomposes full responses into overlapping single-flash responses. These single-flash responses are used to predict responses to novel stimulation sequences, which in turn serve as templates for classification. The linear generative model explains on average 50% and up to 66% of the variance of responses to both seen and unseen sequences. In an online experiment, 12 participants tested a 6 × 6 matrix speller BCI. On average, an online accuracy of 86% was reached with trial lengths of 3.21 seconds. This corresponds to an Information Transfer Rate of 48 bits per minute (approximately 9 symbols per minute). This study indicates the potential to model and predict responses to broad-band stimulation. These predicted responses are proven to be well-suited as templates for a BBVEP-based BCI, thereby enabling communication and control by brain activity only.

No MeSH data available.


The online pipeline.Three stages exist: training, calibration and testing. During training, responses X to stimuli from V are recorded. During calibration, X is deconvolved to pulse responses r using V. Template responses TV and TU are generated by convolving these r with the bit-sequences V and U, respectively. Templates are multiplied (circles) with filters (WX, WT) designed by CCA. The subset and layout are optimized giving U′ and , and stopping margins m are learned. In the testing phase, a new single-trial x is assigned the class-label y that maximizes the correlation between the spatially filtered single-trial x and templates . The classifier emits the class-label if the maximum correlation exceeds the threshold margin. In the case wherein the margin is not reached, more data is collected.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4514763&req=5

pone.0133797.g003: The online pipeline.Three stages exist: training, calibration and testing. During training, responses X to stimuli from V are recorded. During calibration, X is deconvolved to pulse responses r using V. Template responses TV and TU are generated by convolving these r with the bit-sequences V and U, respectively. Templates are multiplied (circles) with filters (WX, WT) designed by CCA. The subset and layout are optimized giving U′ and , and stopping margins m are learned. In the testing phase, a new single-trial x is assigned the class-label y that maximizes the correlation between the spatially filtered single-trial x and templates . The classifier emits the class-label if the maximum correlation exceeds the threshold margin. In the case wherein the margin is not reached, more data is collected.

Mentions: The classifier, including the templates, was constructed following five consecutive steps: template generation, spatial filtering, subset optimization, layout optimization, and learning of stopping criteria (see Fig 3). Matlab routines are available at GitHub (https://github.com/thijor/Reconvolution).


Broad-Band Visually Evoked Potentials: Re(con)volution in Brain-Computer Interfacing.

Thielen J, van den Broek P, Farquhar J, Desain P - PLoS ONE (2015)

The online pipeline.Three stages exist: training, calibration and testing. During training, responses X to stimuli from V are recorded. During calibration, X is deconvolved to pulse responses r using V. Template responses TV and TU are generated by convolving these r with the bit-sequences V and U, respectively. Templates are multiplied (circles) with filters (WX, WT) designed by CCA. The subset and layout are optimized giving U′ and , and stopping margins m are learned. In the testing phase, a new single-trial x is assigned the class-label y that maximizes the correlation between the spatially filtered single-trial x and templates . The classifier emits the class-label if the maximum correlation exceeds the threshold margin. In the case wherein the margin is not reached, more data is collected.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4514763&req=5

pone.0133797.g003: The online pipeline.Three stages exist: training, calibration and testing. During training, responses X to stimuli from V are recorded. During calibration, X is deconvolved to pulse responses r using V. Template responses TV and TU are generated by convolving these r with the bit-sequences V and U, respectively. Templates are multiplied (circles) with filters (WX, WT) designed by CCA. The subset and layout are optimized giving U′ and , and stopping margins m are learned. In the testing phase, a new single-trial x is assigned the class-label y that maximizes the correlation between the spatially filtered single-trial x and templates . The classifier emits the class-label if the maximum correlation exceeds the threshold margin. In the case wherein the margin is not reached, more data is collected.
Mentions: The classifier, including the templates, was constructed following five consecutive steps: template generation, spatial filtering, subset optimization, layout optimization, and learning of stopping criteria (see Fig 3). Matlab routines are available at GitHub (https://github.com/thijor/Reconvolution).

Bottom Line: We defined a linear generative model that decomposes full responses into overlapping single-flash responses.In an online experiment, 12 participants tested a 6 × 6 matrix speller BCI.These predicted responses are proven to be well-suited as templates for a BBVEP-based BCI, thereby enabling communication and control by brain activity only.

View Article: PubMed Central - PubMed

Affiliation: Radboud University Nijmegen, Donders Center for Cognition, Nijmegen, Netherlands.

ABSTRACT
Brain-Computer Interfaces (BCIs) allow users to control devices and communicate by using brain activity only. BCIs based on broad-band visual stimulation can outperform BCIs using other stimulation paradigms. Visual stimulation with pseudo-random bit-sequences evokes specific Broad-Band Visually Evoked Potentials (BBVEPs) that can be reliably used in BCI for high-speed communication in speller applications. In this study, we report a novel paradigm for a BBVEP-based BCI that utilizes a generative framework to predict responses to broad-band stimulation sequences. In this study we designed a BBVEP-based BCI using modulated Gold codes to mark cells in a visual speller BCI. We defined a linear generative model that decomposes full responses into overlapping single-flash responses. These single-flash responses are used to predict responses to novel stimulation sequences, which in turn serve as templates for classification. The linear generative model explains on average 50% and up to 66% of the variance of responses to both seen and unseen sequences. In an online experiment, 12 participants tested a 6 × 6 matrix speller BCI. On average, an online accuracy of 86% was reached with trial lengths of 3.21 seconds. This corresponds to an Information Transfer Rate of 48 bits per minute (approximately 9 symbols per minute). This study indicates the potential to model and predict responses to broad-band stimulation. These predicted responses are proven to be well-suited as templates for a BBVEP-based BCI, thereby enabling communication and control by brain activity only.

No MeSH data available.